TRACK: Multiple Instance Learning for Anti-Money Laundering in Transaction Networks (arXiv (Cornell University))

Summary

• The paper recasts account-level anti-money laundering detection as multiple instance learning, with an account-window as a bag and its transactions as instances.p.1
• It names the discrepancy between account-level labels and aggregated transaction summaries the aggregation mismatch.p.1
• Each transaction instance has six transaction attributes and 22 history features for the focal account, up to 50 in total.p.4
• Evaluation uses four public benchmarks and one non-public real-world interbank network.p.5
• Transactions are ordered by time and split 60/20/20 into training, validation and test sets.p.5
• A post-hoc estimate indicates that 96.9–100.0% of positive accounts retain at least one flagged transaction.p.10

Key insights

1. Aggregating many transactions can dilute the signal and obscure which transactions support an account-level prediction.p.2
2. TRACK leads every baseline trained with account-level supervision across all five datasets.p.6
3. TRACK yields account risk and transaction ranking from account labels alone.p.2
4. Median bag sizes are 2 transactions on HOFINET and Elliptic++, 6 on AMLworld, 189 on AMLNet and 216 on AMLSim.p.10
5. Selective aggregation after instance encoding preserves evidence that pre-aggregation averaging loses.p.9